2 % (c) The University of Glasgow 2002-2006
5 ByteCodeGen: Generate bytecode from Core
8 module ByteCodeGen ( UnlinkedBCO, byteCodeGen, coreExprToBCOs ) where
10 #include "HsVersions.h"
50 import Data.List ( intersperse, sortBy, zip4, zip6, partition )
51 import Foreign ( Ptr, castPtr, mallocBytes, pokeByteOff, Word8,
52 withForeignPtr, castFunPtrToPtr, nullPtr, plusPtr )
54 import Control.Exception ( throwDyn )
56 import GHC.Exts ( Int(..), ByteArray# )
58 import Control.Monad ( when )
59 import Data.Char ( ord, chr )
67 -- -----------------------------------------------------------------------------
68 -- Generating byte code for a complete module
70 byteCodeGen :: DynFlags
74 -> IO CompiledByteCode
75 byteCodeGen dflags binds tycs modBreaks
76 = do showPass dflags "ByteCodeGen"
78 let flatBinds = [ (bndr, freeVars rhs)
79 | (bndr, rhs) <- flattenBinds binds]
81 us <- mkSplitUniqSupply 'y'
82 (BcM_State _us final_ctr mallocd _, proto_bcos)
83 <- runBc us modBreaks (mapM schemeTopBind flatBinds)
85 when (notNull mallocd)
86 (panic "ByteCodeGen.byteCodeGen: missing final emitBc?")
88 dumpIfSet_dyn dflags Opt_D_dump_BCOs
89 "Proto-BCOs" (vcat (intersperse (char ' ') (map ppr proto_bcos)))
91 assembleBCOs proto_bcos tycs
93 -- -----------------------------------------------------------------------------
94 -- Generating byte code for an expression
96 -- Returns: (the root BCO for this expression,
97 -- a list of auxilary BCOs resulting from compiling closures)
98 coreExprToBCOs :: DynFlags
101 coreExprToBCOs dflags expr
102 = do showPass dflags "ByteCodeGen"
104 -- create a totally bogus name for the top-level BCO; this
105 -- should be harmless, since it's never used for anything
106 let invented_name = mkSystemVarName (mkPseudoUniqueE 0) FSLIT("ExprTopLevel")
107 invented_id = Id.mkLocalId invented_name (panic "invented_id's type")
109 -- the uniques are needed to generate fresh variables when we introduce new
110 -- let bindings for ticked expressions
111 us <- mkSplitUniqSupply 'y'
112 (BcM_State _us final_ctr mallocd _ , proto_bco)
113 <- runBc us emptyModBreaks (schemeTopBind (invented_id, freeVars expr))
115 when (notNull mallocd)
116 (panic "ByteCodeGen.coreExprToBCOs: missing final emitBc?")
118 dumpIfSet_dyn dflags Opt_D_dump_BCOs "Proto-BCOs" (ppr proto_bco)
120 assembleBCO proto_bco
123 -- -----------------------------------------------------------------------------
124 -- Compilation schema for the bytecode generator
126 type BCInstrList = OrdList BCInstr
128 type Sequel = Int -- back off to this depth before ENTER
130 -- Maps Ids to the offset from the stack _base_ so we don't have
131 -- to mess with it after each push/pop.
132 type BCEnv = FiniteMap Id Int -- To find vars on the stack
134 ppBCEnv :: BCEnv -> SDoc
137 $$ nest 4 (vcat (map pp_one (sortBy cmp_snd (fmToList p))))
140 pp_one (var, offset) = int offset <> colon <+> ppr var <+> ppr (idCgRep var)
141 cmp_snd x y = compare (snd x) (snd y)
143 -- Create a BCO and do a spot of peephole optimisation on the insns
148 -> Either [AnnAlt Id VarSet] (AnnExpr Id VarSet)
152 -> Bool -- True <=> is a return point, rather than a function
155 mkProtoBCO nm instrs_ordlist origin arity bitmap_size bitmap is_ret mallocd_blocks
158 protoBCOInstrs = maybe_with_stack_check,
159 protoBCOBitmap = bitmap,
160 protoBCOBitmapSize = bitmap_size,
161 protoBCOArity = arity,
162 protoBCOExpr = origin,
163 protoBCOPtrs = mallocd_blocks
166 -- Overestimate the stack usage (in words) of this BCO,
167 -- and if >= iNTERP_STACK_CHECK_THRESH, add an explicit
168 -- stack check. (The interpreter always does a stack check
169 -- for iNTERP_STACK_CHECK_THRESH words at the start of each
170 -- BCO anyway, so we only need to add an explicit on in the
171 -- (hopefully rare) cases when the (overestimated) stack use
172 -- exceeds iNTERP_STACK_CHECK_THRESH.
173 maybe_with_stack_check
175 -- don't do stack checks at return points;
176 -- everything is aggregated up to the top BCO
177 -- (which must be a function)
178 | stack_overest >= iNTERP_STACK_CHECK_THRESH
179 = STKCHECK stack_overest : peep_d
181 = peep_d -- the supposedly common case
183 -- We assume that this sum doesn't wrap
184 stack_overest = sum (map bciStackUse peep_d)
186 -- Merge local pushes
187 peep_d = peep (fromOL instrs_ordlist)
189 peep (PUSH_L off1 : PUSH_L off2 : PUSH_L off3 : rest)
190 = PUSH_LLL off1 (off2-1) (off3-2) : peep rest
191 peep (PUSH_L off1 : PUSH_L off2 : rest)
192 = PUSH_LL off1 (off2-1) : peep rest
198 argBits :: [CgRep] -> [Bool]
201 | isFollowableArg rep = False : argBits args
202 | otherwise = take (cgRepSizeW rep) (repeat True) ++ argBits args
204 -- -----------------------------------------------------------------------------
207 -- Compile code for the right-hand side of a top-level binding
209 schemeTopBind :: (Id, AnnExpr Id VarSet) -> BcM (ProtoBCO Name)
212 schemeTopBind (id, rhs)
213 | Just data_con <- isDataConWorkId_maybe id,
214 isNullaryRepDataCon data_con = do
215 -- Special case for the worker of a nullary data con.
216 -- It'll look like this: Nil = /\a -> Nil a
217 -- If we feed it into schemeR, we'll get
219 -- because mkConAppCode treats nullary constructor applications
220 -- by just re-using the single top-level definition. So
221 -- for the worker itself, we must allocate it directly.
222 -- ioToBc (putStrLn $ "top level BCO")
223 emitBc (mkProtoBCO (getName id) (toOL [PACK data_con 0, ENTER])
224 (Right rhs) 0 0 [{-no bitmap-}] False{-not alts-})
227 = schemeR [{- No free variables -}] (id, rhs)
230 -- -----------------------------------------------------------------------------
233 -- Compile code for a right-hand side, to give a BCO that,
234 -- when executed with the free variables and arguments on top of the stack,
235 -- will return with a pointer to the result on top of the stack, after
236 -- removing the free variables and arguments.
238 -- Park the resulting BCO in the monad. Also requires the
239 -- variable to which this value was bound, so as to give the
240 -- resulting BCO a name.
242 schemeR :: [Id] -- Free vars of the RHS, ordered as they
243 -- will appear in the thunk. Empty for
244 -- top-level things, which have no free vars.
245 -> (Id, AnnExpr Id VarSet)
246 -> BcM (ProtoBCO Name)
247 schemeR fvs (nm, rhs)
251 $$ (ppr.filter (not.isTyVar).varSetElems.fst) rhs
252 $$ pprCoreExpr (deAnnotate rhs)
258 = schemeR_wrk fvs nm rhs (collect [] rhs)
260 collect :: [Var] -> AnnExpr Id VarSet -> ([Var], AnnExpr' Id VarSet)
261 collect xs (_, AnnNote note e) = collect xs e
262 collect xs (_, AnnCast e _) = collect xs e
263 collect xs (_, AnnLam x e) = collect (if isTyVar x then xs else (x:xs)) e
264 collect xs (_, not_lambda) = (reverse xs, not_lambda)
266 schemeR_wrk :: [Id] -> Id -> AnnExpr Id VarSet -> ([Var], AnnExpr' Var VarSet) -> BcM (ProtoBCO Name)
267 schemeR_wrk fvs nm original_body (args, body)
269 all_args = reverse args ++ fvs
270 arity = length all_args
271 -- all_args are the args in reverse order. We're compiling a function
272 -- \fv1..fvn x1..xn -> e
273 -- i.e. the fvs come first
275 szsw_args = map idSizeW all_args
276 szw_args = sum szsw_args
277 p_init = listToFM (zip all_args (mkStackOffsets 0 szsw_args))
279 -- make the arg bitmap
280 bits = argBits (reverse (map idCgRep all_args))
281 bitmap_size = length bits
282 bitmap = mkBitmap bits
284 body_code <- schemeER_wrk szw_args p_init body
286 emitBc (mkProtoBCO (getName nm) body_code (Right original_body)
287 arity bitmap_size bitmap False{-not alts-})
289 -- introduce break instructions for ticked expressions
290 schemeER_wrk :: Int -> BCEnv -> AnnExpr' Id VarSet -> BcM BCInstrList
292 | Just (tickInfo, (_annot, newRhs)) <- isTickedExp' rhs = do
293 code <- schemeE d 0 p newRhs
295 let idOffSets = getVarOffSets d p tickInfo
296 let tickNumber = tickInfo_number tickInfo
297 let breakInfo = BreakInfo
298 { breakInfo_module = tickInfo_module tickInfo
299 , breakInfo_number = tickNumber
300 , breakInfo_vars = idOffSets
302 let breakInstr = case arr of (BA arr#) -> BRK_FUN arr# tickNumber breakInfo
303 return $ breakInstr `consOL` code
304 | otherwise = schemeE d 0 p rhs
306 getVarOffSets :: Int -> BCEnv -> TickInfo -> [(Id, Int)]
307 getVarOffSets d p = catMaybes . map (getOffSet d p) . tickInfo_locals
309 getOffSet :: Int -> BCEnv -> Id -> Maybe (Id, Int)
311 = case lookupBCEnv_maybe env id of
313 Just offset -> Just (id, d - offset)
315 fvsToEnv :: BCEnv -> VarSet -> [Id]
316 -- Takes the free variables of a right-hand side, and
317 -- delivers an ordered list of the local variables that will
318 -- be captured in the thunk for the RHS
319 -- The BCEnv argument tells which variables are in the local
320 -- environment: these are the ones that should be captured
322 -- The code that constructs the thunk, and the code that executes
323 -- it, have to agree about this layout
324 fvsToEnv p fvs = [v | v <- varSetElems fvs,
325 isId v, -- Could be a type variable
328 -- -----------------------------------------------------------------------------
333 { tickInfo_number :: Int -- the (module) unique number of the tick
334 , tickInfo_module :: Module -- the origin of the ticked expression
335 , tickInfo_locals :: [Id] -- the local vars in scope at the ticked expression
338 instance Outputable TickInfo where
339 ppr info = text "TickInfo" <+>
340 parens (int (tickInfo_number info) <+> ppr (tickInfo_module info) <+>
341 ppr (tickInfo_locals info))
343 -- Compile code to apply the given expression to the remaining args
344 -- on the stack, returning a HNF.
345 schemeE :: Int -> Sequel -> BCEnv -> AnnExpr' Id VarSet -> BcM BCInstrList
347 -- Delegate tail-calls to schemeT.
348 schemeE d s p e@(AnnApp f a)
351 schemeE d s p e@(AnnVar v)
352 | not (isUnLiftedType v_type)
353 = -- Lifted-type thing; push it in the normal way
357 = do -- Returning an unlifted value.
358 -- Heave it on the stack, SLIDE, and RETURN.
359 (push, szw) <- pushAtom d p (AnnVar v)
360 return (push -- value onto stack
361 `appOL` mkSLIDE szw (d-s) -- clear to sequel
362 `snocOL` RETURN_UBX v_rep) -- go
365 v_rep = typeCgRep v_type
367 schemeE d s p (AnnLit literal)
368 = do (push, szw) <- pushAtom d p (AnnLit literal)
369 let l_rep = typeCgRep (literalType literal)
370 return (push -- value onto stack
371 `appOL` mkSLIDE szw (d-s) -- clear to sequel
372 `snocOL` RETURN_UBX l_rep) -- go
374 schemeE d s p (AnnLet (AnnNonRec x (_,rhs)) (_,body))
375 | (AnnVar v, args_r_to_l) <- splitApp rhs,
376 Just data_con <- isDataConWorkId_maybe v,
377 dataConRepArity data_con == length args_r_to_l
378 = do -- Special case for a non-recursive let whose RHS is a
379 -- saturatred constructor application.
380 -- Just allocate the constructor and carry on
381 alloc_code <- mkConAppCode d s p data_con args_r_to_l
382 body_code <- schemeE (d+1) s (addToFM p x d) body
383 return (alloc_code `appOL` body_code)
385 -- General case for let. Generates correct, if inefficient, code in
387 schemeE d s p (AnnLet binds (_,body))
388 = let (xs,rhss) = case binds of AnnNonRec x rhs -> ([x],[rhs])
389 AnnRec xs_n_rhss -> unzip xs_n_rhss
392 fvss = map (fvsToEnv p' . fst) rhss
394 -- Sizes of free vars
395 sizes = map (\rhs_fvs -> sum (map idSizeW rhs_fvs)) fvss
397 -- the arity of each rhs
398 arities = map (length . fst . collect []) rhss
400 -- This p', d' defn is safe because all the items being pushed
401 -- are ptrs, so all have size 1. d' and p' reflect the stack
402 -- after the closures have been allocated in the heap (but not
403 -- filled in), and pointers to them parked on the stack.
404 p' = addListToFM p (zipE xs (mkStackOffsets d (nOfThem n_binds 1)))
406 zipE = zipEqual "schemeE"
408 -- ToDo: don't build thunks for things with no free variables
409 build_thunk dd [] size bco off arity
410 = return (PUSH_BCO bco `consOL` unitOL (mkap (off+size) size))
412 mkap | arity == 0 = MKAP
414 build_thunk dd (fv:fvs) size bco off arity = do
415 (push_code, pushed_szw) <- pushAtom dd p' (AnnVar fv)
416 more_push_code <- build_thunk (dd+pushed_szw) fvs size bco off arity
417 return (push_code `appOL` more_push_code)
419 alloc_code = toOL (zipWith mkAlloc sizes arities)
420 where mkAlloc sz 0 = ALLOC_AP sz
421 mkAlloc sz arity = ALLOC_PAP arity sz
423 compile_bind d' fvs x rhs size arity off = do
424 bco <- schemeR fvs (x,rhs)
425 build_thunk d' fvs size bco off arity
428 [ compile_bind d' fvs x rhs size arity n
429 | (fvs, x, rhs, size, arity, n) <-
430 zip6 fvss xs rhss sizes arities [n_binds, n_binds-1 .. 1]
433 body_code <- schemeE d' s p' body
434 thunk_codes <- sequence compile_binds
435 return (alloc_code `appOL` concatOL thunk_codes `appOL` body_code)
437 -- introduce a let binding for a ticked case expression. This rule *should* only fire when the
438 -- expression was not already let-bound (the code gen for let bindings should take care of that).
439 -- Todo: we call exprFreeVars on a deAnnotated expression, this may not be the best way
440 -- to calculate the free vars but it seemed like the least intrusive thing to do
441 schemeE d s p exp@(AnnCase {})
442 | Just (tickInfo, _exp) <- isTickedExp' exp = do
443 let fvs = exprFreeVars $ deAnnotate' exp
444 let ty = exprType $ deAnnotate' exp
446 -- Todo: is emptyVarSet correct on the next line?
447 let letExp = AnnLet (AnnNonRec id (fvs, exp)) (emptyVarSet, AnnVar id)
450 schemeE d s p (AnnCase scrut bndr _ [(DataAlt dc, [bind1, bind2], rhs)])
451 | isUnboxedTupleCon dc, VoidArg <- typeCgRep (idType bind1)
453 -- case .... of x { (# VoidArg'd-thing, a #) -> ... }
455 -- case .... of a { DEFAULT -> ... }
456 -- becuse the return convention for both are identical.
458 -- Note that it does not matter losing the void-rep thing from the
459 -- envt (it won't be bound now) because we never look such things up.
461 = --trace "automagic mashing of case alts (# VoidArg, a #)" $
462 doCase d s p scrut bind2 [(DEFAULT, [], rhs)] True{-unboxed tuple-}
464 | isUnboxedTupleCon dc, VoidArg <- typeCgRep (idType bind2)
465 = --trace "automagic mashing of case alts (# a, VoidArg #)" $
466 doCase d s p scrut bind1 [(DEFAULT, [], rhs)] True{-unboxed tuple-}
468 schemeE d s p (AnnCase scrut bndr _ [(DataAlt dc, [bind1], rhs)])
469 | isUnboxedTupleCon dc
470 -- Similarly, convert
471 -- case .... of x { (# a #) -> ... }
473 -- case .... of a { DEFAULT -> ... }
474 = --trace "automagic mashing of case alts (# a #)" $
475 doCase d s p scrut bind1 [(DEFAULT, [], rhs)] True{-unboxed tuple-}
477 schemeE d s p (AnnCase scrut bndr _ alts)
478 = doCase d s p scrut bndr alts False{-not an unboxed tuple-}
480 schemeE d s p (AnnNote note (_, body))
483 schemeE d s p (AnnCast (_, body) _)
487 = pprPanic "ByteCodeGen.schemeE: unhandled case"
488 (pprCoreExpr (deAnnotate' other))
494 A ticked expression looks like this:
496 case tick<n> var1 ... varN of DEFAULT -> e
498 (*) <n> is the number of the tick, which is unique within a module
499 (*) var1 ... varN are the local variables in scope at the tick site
501 If we find a ticked expression we return:
503 Just ((n, [var1 ... varN]), e)
505 otherwise we return Nothing.
507 The idea is that the "case tick<n> ..." is really just an annotation on
508 the code. When we find such a thing, we pull out the useful information,
509 and then compile the code as if it was just the expression "e".
513 isTickedExp :: AnnExpr Id a -> Maybe (TickInfo, AnnExpr Id a)
514 isTickedExp (annot, expr) = isTickedExp' expr
516 isTickedExp' :: AnnExpr' Id a -> Maybe (TickInfo, AnnExpr Id a)
517 isTickedExp' (AnnCase scrut _bndr _type alts)
518 | Just tickInfo <- isTickedScrut scrut,
519 [(DEFAULT, _bndr, rhs)] <- alts
520 = Just (tickInfo, rhs)
522 isTickedScrut :: (AnnExpr Id a) -> Maybe TickInfo
525 Just (TickBox modName tickNumber) <- isTickBoxOp_maybe id
526 = Just $ TickInfo { tickInfo_number = tickNumber
527 , tickInfo_module = modName
528 , tickInfo_locals = idsOfArgs args
530 | otherwise = Nothing
532 (f, args) = collectArgs $ deAnnotate expr
533 idsOfArgs :: [Expr Id] -> [Id]
534 idsOfArgs = catMaybes . map exprId
535 exprId :: Expr Id -> Maybe Id
536 exprId (Var id) = Just id
537 exprId other = Nothing
539 isTickedExp' other = Nothing
541 -- Compile code to do a tail call. Specifically, push the fn,
542 -- slide the on-stack app back down to the sequel depth,
543 -- and enter. Four cases:
546 -- An application "GHC.Prim.tagToEnum# <type> unboxed-int".
547 -- The int will be on the stack. Generate a code sequence
548 -- to convert it to the relevant constructor, SLIDE and ENTER.
550 -- 1. The fn denotes a ccall. Defer to generateCCall.
552 -- 2. (Another nasty hack). Spot (# a::VoidArg, b #) and treat
553 -- it simply as b -- since the representations are identical
554 -- (the VoidArg takes up zero stack space). Also, spot
555 -- (# b #) and treat it as b.
557 -- 3. Application of a constructor, by defn saturated.
558 -- Split the args into ptrs and non-ptrs, and push the nonptrs,
559 -- then the ptrs, and then do PACK and RETURN.
561 -- 4. Otherwise, it must be a function call. Push the args
562 -- right to left, SLIDE and ENTER.
564 schemeT :: Int -- Stack depth
565 -> Sequel -- Sequel depth
566 -> BCEnv -- stack env
567 -> AnnExpr' Id VarSet
572 -- | trace ("schemeT: env in = \n" ++ showSDocDebug (ppBCEnv p)) False
573 -- = panic "schemeT ?!?!"
575 -- | trace ("\nschemeT\n" ++ showSDoc (pprCoreExpr (deAnnotate' app)) ++ "\n") False
579 | Just (arg, constr_names) <- maybe_is_tagToEnum_call
580 = do (push, arg_words) <- pushAtom d p arg
581 tagToId_sequence <- implement_tagToId constr_names
582 return (push `appOL` tagToId_sequence
583 `appOL` mkSLIDE 1 (d+arg_words-s)
587 | Just (CCall ccall_spec) <- isFCallId_maybe fn
588 = generateCCall d s p ccall_spec fn args_r_to_l
590 -- Case 2: Constructor application
591 | Just con <- maybe_saturated_dcon,
592 isUnboxedTupleCon con
593 = case args_r_to_l of
594 [arg1,arg2] | isVoidArgAtom arg1 ->
595 unboxedTupleReturn d s p arg2
596 [arg1,arg2] | isVoidArgAtom arg2 ->
597 unboxedTupleReturn d s p arg1
598 _other -> unboxedTupleException
600 -- Case 3: Ordinary data constructor
601 | Just con <- maybe_saturated_dcon
602 = do alloc_con <- mkConAppCode d s p con args_r_to_l
603 return (alloc_con `appOL`
604 mkSLIDE 1 (d - s) `snocOL`
607 -- Case 4: Tail call of function
609 = doTailCall d s p fn args_r_to_l
612 -- Detect and extract relevant info for the tagToEnum kludge.
613 maybe_is_tagToEnum_call
614 = let extract_constr_Names ty
615 | Just (tyc, []) <- splitTyConApp_maybe (repType ty),
617 = map (getName . dataConWorkId) (tyConDataCons tyc)
618 -- NOTE: use the worker name, not the source name of
619 -- the DataCon. See DataCon.lhs for details.
621 = panic "maybe_is_tagToEnum_call.extract_constr_Ids"
624 (AnnApp (_, AnnApp (_, AnnVar v) (_, AnnType t)) arg)
625 -> case isPrimOpId_maybe v of
626 Just TagToEnumOp -> Just (snd arg, extract_constr_Names t)
630 -- Extract the args (R->L) and fn
631 -- The function will necessarily be a variable,
632 -- because we are compiling a tail call
633 (AnnVar fn, args_r_to_l) = splitApp app
635 -- Only consider this to be a constructor application iff it is
636 -- saturated. Otherwise, we'll call the constructor wrapper.
637 n_args = length args_r_to_l
639 = case isDataConWorkId_maybe fn of
640 Just con | dataConRepArity con == n_args -> Just con
643 -- -----------------------------------------------------------------------------
644 -- Generate code to build a constructor application,
645 -- leaving it on top of the stack
647 mkConAppCode :: Int -> Sequel -> BCEnv
648 -> DataCon -- The data constructor
649 -> [AnnExpr' Id VarSet] -- Args, in *reverse* order
652 mkConAppCode orig_d s p con [] -- Nullary constructor
653 = ASSERT( isNullaryRepDataCon con )
654 return (unitOL (PUSH_G (getName (dataConWorkId con))))
655 -- Instead of doing a PACK, which would allocate a fresh
656 -- copy of this constructor, use the single shared version.
658 mkConAppCode orig_d s p con args_r_to_l
659 = ASSERT( dataConRepArity con == length args_r_to_l )
660 do_pushery orig_d (non_ptr_args ++ ptr_args)
662 -- The args are already in reverse order, which is the way PACK
663 -- expects them to be. We must push the non-ptrs after the ptrs.
664 (ptr_args, non_ptr_args) = partition isPtrAtom args_r_to_l
666 do_pushery d (arg:args)
667 = do (push, arg_words) <- pushAtom d p arg
668 more_push_code <- do_pushery (d+arg_words) args
669 return (push `appOL` more_push_code)
671 = return (unitOL (PACK con n_arg_words))
673 n_arg_words = d - orig_d
676 -- -----------------------------------------------------------------------------
677 -- Returning an unboxed tuple with one non-void component (the only
678 -- case we can handle).
680 -- Remember, we don't want to *evaluate* the component that is being
681 -- returned, even if it is a pointed type. We always just return.
684 :: Int -> Sequel -> BCEnv
685 -> AnnExpr' Id VarSet -> BcM BCInstrList
686 unboxedTupleReturn d s p arg = do
687 (push, sz) <- pushAtom d p arg
689 mkSLIDE sz (d-s) `snocOL`
690 RETURN_UBX (atomRep arg))
692 -- -----------------------------------------------------------------------------
693 -- Generate code for a tail-call
696 :: Int -> Sequel -> BCEnv
697 -> Id -> [AnnExpr' Id VarSet]
699 doTailCall init_d s p fn args
700 = do_pushes init_d args (map atomRep args)
702 do_pushes d [] reps = do
703 ASSERT( null reps ) return ()
704 (push_fn, sz) <- pushAtom d p (AnnVar fn)
705 ASSERT( sz == 1 ) return ()
706 return (push_fn `appOL` (
707 mkSLIDE ((d-init_d) + 1) (init_d - s) `appOL`
709 do_pushes d args reps = do
710 let (push_apply, n, rest_of_reps) = findPushSeq reps
711 (these_args, rest_of_args) = splitAt n args
712 (next_d, push_code) <- push_seq d these_args
713 instrs <- do_pushes (next_d + 1) rest_of_args rest_of_reps
714 -- ^^^ for the PUSH_APPLY_ instruction
715 return (push_code `appOL` (push_apply `consOL` instrs))
717 push_seq d [] = return (d, nilOL)
718 push_seq d (arg:args) = do
719 (push_code, sz) <- pushAtom d p arg
720 (final_d, more_push_code) <- push_seq (d+sz) args
721 return (final_d, push_code `appOL` more_push_code)
723 -- v. similar to CgStackery.findMatch, ToDo: merge
724 findPushSeq (PtrArg: PtrArg: PtrArg: PtrArg: PtrArg: PtrArg: rest)
725 = (PUSH_APPLY_PPPPPP, 6, rest)
726 findPushSeq (PtrArg: PtrArg: PtrArg: PtrArg: PtrArg: rest)
727 = (PUSH_APPLY_PPPPP, 5, rest)
728 findPushSeq (PtrArg: PtrArg: PtrArg: PtrArg: rest)
729 = (PUSH_APPLY_PPPP, 4, rest)
730 findPushSeq (PtrArg: PtrArg: PtrArg: rest)
731 = (PUSH_APPLY_PPP, 3, rest)
732 findPushSeq (PtrArg: PtrArg: rest)
733 = (PUSH_APPLY_PP, 2, rest)
734 findPushSeq (PtrArg: rest)
735 = (PUSH_APPLY_P, 1, rest)
736 findPushSeq (VoidArg: rest)
737 = (PUSH_APPLY_V, 1, rest)
738 findPushSeq (NonPtrArg: rest)
739 = (PUSH_APPLY_N, 1, rest)
740 findPushSeq (FloatArg: rest)
741 = (PUSH_APPLY_F, 1, rest)
742 findPushSeq (DoubleArg: rest)
743 = (PUSH_APPLY_D, 1, rest)
744 findPushSeq (LongArg: rest)
745 = (PUSH_APPLY_L, 1, rest)
747 = panic "ByteCodeGen.findPushSeq"
749 -- -----------------------------------------------------------------------------
752 doCase :: Int -> Sequel -> BCEnv
753 -> AnnExpr Id VarSet -> Id -> [AnnAlt Id VarSet]
754 -> Bool -- True <=> is an unboxed tuple case, don't enter the result
756 doCase d s p (_,scrut) bndr alts is_unboxed_tuple
758 -- Top of stack is the return itbl, as usual.
759 -- underneath it is the pointer to the alt_code BCO.
760 -- When an alt is entered, it assumes the returned value is
761 -- on top of the itbl.
764 -- An unlifted value gets an extra info table pushed on top
765 -- when it is returned.
766 unlifted_itbl_sizeW | isAlgCase = 0
769 -- depth of stack after the return value has been pushed
770 d_bndr = d + ret_frame_sizeW + idSizeW bndr
772 -- depth of stack after the extra info table for an unboxed return
773 -- has been pushed, if any. This is the stack depth at the
775 d_alts = d_bndr + unlifted_itbl_sizeW
777 -- Env in which to compile the alts, not including
778 -- any vars bound by the alts themselves
779 p_alts = addToFM p bndr (d_bndr - 1)
781 bndr_ty = idType bndr
782 isAlgCase = not (isUnLiftedType bndr_ty) && not is_unboxed_tuple
784 -- given an alt, return a discr and code for it.
785 codeAlt alt@(DEFAULT, _, (_,rhs))
786 = do rhs_code <- schemeE d_alts s p_alts rhs
787 return (NoDiscr, rhs_code)
789 codeAlt alt@(discr, bndrs, (_,rhs))
790 -- primitive or nullary constructor alt: no need to UNPACK
791 | null real_bndrs = do
792 rhs_code <- schemeE d_alts s p_alts rhs
793 return (my_discr alt, rhs_code)
794 -- algebraic alt with some binders
795 | ASSERT(isAlgCase) otherwise =
797 (ptrs,nptrs) = partition (isFollowableArg.idCgRep) real_bndrs
798 ptr_sizes = map idSizeW ptrs
799 nptrs_sizes = map idSizeW nptrs
800 bind_sizes = ptr_sizes ++ nptrs_sizes
801 size = sum ptr_sizes + sum nptrs_sizes
802 -- the UNPACK instruction unpacks in reverse order...
803 p' = addListToFM p_alts
804 (zip (reverse (ptrs ++ nptrs))
805 (mkStackOffsets d_alts (reverse bind_sizes)))
807 rhs_code <- schemeE (d_alts+size) s p' rhs
808 return (my_discr alt, unitOL (UNPACK size) `appOL` rhs_code)
810 real_bndrs = filter (not.isTyVar) bndrs
812 my_discr (DEFAULT, binds, rhs) = NoDiscr {-shouldn't really happen-}
813 my_discr (DataAlt dc, binds, rhs)
814 | isUnboxedTupleCon dc
815 = unboxedTupleException
817 = DiscrP (dataConTag dc - fIRST_TAG)
818 my_discr (LitAlt l, binds, rhs)
819 = case l of MachInt i -> DiscrI (fromInteger i)
820 MachFloat r -> DiscrF (fromRational r)
821 MachDouble r -> DiscrD (fromRational r)
822 MachChar i -> DiscrI (ord i)
823 _ -> pprPanic "schemeE(AnnCase).my_discr" (ppr l)
826 | not isAlgCase = Nothing
828 = case [dc | (DataAlt dc, _, _) <- alts] of
830 (dc:_) -> Just (tyConFamilySize (dataConTyCon dc))
832 -- the bitmap is relative to stack depth d, i.e. before the
833 -- BCO, info table and return value are pushed on.
834 -- This bit of code is v. similar to buildLivenessMask in CgBindery,
835 -- except that here we build the bitmap from the known bindings of
836 -- things that are pointers, whereas in CgBindery the code builds the
837 -- bitmap from the free slots and unboxed bindings.
840 -- NOTE [7/12/2006] bug #1013, testcase ghci/should_run/ghci002.
841 -- The bitmap must cover the portion of the stack up to the sequel only.
842 -- Previously we were building a bitmap for the whole depth (d), but we
843 -- really want a bitmap up to depth (d-s). This affects compilation of
844 -- case-of-case expressions, which is the only time we can be compiling a
845 -- case expression with s /= 0.
847 bitmap = intsToReverseBitmap bitmap_size{-size-}
848 (sortLe (<=) (filter (< bitmap_size) rel_slots))
851 rel_slots = concat (map spread binds)
853 | isFollowableArg (idCgRep id) = [ rel_offset ]
855 where rel_offset = d - offset - 1
858 alt_stuff <- mapM codeAlt alts
859 alt_final <- mkMultiBranch maybe_ncons alt_stuff
862 alt_bco_name = getName bndr
863 alt_bco = mkProtoBCO alt_bco_name alt_final (Left alts)
864 0{-no arity-} bitmap_size bitmap True{-is alts-}
866 -- trace ("case: bndr = " ++ showSDocDebug (ppr bndr) ++ "\ndepth = " ++ show d ++ "\nenv = \n" ++ showSDocDebug (ppBCEnv p) ++
867 -- "\n bitmap = " ++ show bitmap) $ do
868 scrut_code <- schemeE (d + ret_frame_sizeW) (d + ret_frame_sizeW) p scrut
869 alt_bco' <- emitBc alt_bco
871 | isAlgCase = PUSH_ALTS alt_bco'
872 | otherwise = PUSH_ALTS_UNLIFTED alt_bco' (typeCgRep bndr_ty)
873 return (push_alts `consOL` scrut_code)
876 -- -----------------------------------------------------------------------------
877 -- Deal with a CCall.
879 -- Taggedly push the args onto the stack R->L,
880 -- deferencing ForeignObj#s and adjusting addrs to point to
881 -- payloads in Ptr/Byte arrays. Then, generate the marshalling
882 -- (machine) code for the ccall, and create bytecodes to call that and
883 -- then return in the right way.
885 generateCCall :: Int -> Sequel -- stack and sequel depths
887 -> CCallSpec -- where to call
888 -> Id -- of target, for type info
889 -> [AnnExpr' Id VarSet] -- args (atoms)
892 generateCCall d0 s p ccall_spec@(CCallSpec target cconv safety) fn args_r_to_l
895 addr_sizeW = cgRepSizeW NonPtrArg
897 -- Get the args on the stack, with tags and suitably
898 -- dereferenced for the CCall. For each arg, return the
899 -- depth to the first word of the bits for that arg, and the
900 -- CgRep of what was actually pushed.
902 pargs d [] = return []
904 = let arg_ty = repType (exprType (deAnnotate' a))
906 in case splitTyConApp_maybe arg_ty of
907 -- Don't push the FO; instead push the Addr# it
910 | t == arrayPrimTyCon || t == mutableArrayPrimTyCon
911 -> do rest <- pargs (d + addr_sizeW) az
912 code <- parg_ArrayishRep arrPtrsHdrSize d p a
913 return ((code,NonPtrArg):rest)
915 | t == byteArrayPrimTyCon || t == mutableByteArrayPrimTyCon
916 -> do rest <- pargs (d + addr_sizeW) az
917 code <- parg_ArrayishRep arrWordsHdrSize d p a
918 return ((code,NonPtrArg):rest)
920 -- Default case: push taggedly, but otherwise intact.
922 -> do (code_a, sz_a) <- pushAtom d p a
923 rest <- pargs (d+sz_a) az
924 return ((code_a, atomRep a) : rest)
926 -- Do magic for Ptr/Byte arrays. Push a ptr to the array on
927 -- the stack but then advance it over the headers, so as to
928 -- point to the payload.
929 parg_ArrayishRep hdrSize d p a
930 = do (push_fo, _) <- pushAtom d p a
931 -- The ptr points at the header. Advance it over the
932 -- header and then pretend this is an Addr#.
933 return (push_fo `snocOL` SWIZZLE 0 hdrSize)
936 code_n_reps <- pargs d0 args_r_to_l
938 (pushs_arg, a_reps_pushed_r_to_l) = unzip code_n_reps
940 push_args = concatOL pushs_arg
941 d_after_args = d0 + sum (map cgRepSizeW a_reps_pushed_r_to_l)
943 | null a_reps_pushed_r_to_l || head a_reps_pushed_r_to_l /= VoidArg
944 = panic "ByteCodeGen.generateCCall: missing or invalid World token?"
946 = reverse (tail a_reps_pushed_r_to_l)
948 -- Now: a_reps_pushed_RAW are the reps which are actually on the stack.
949 -- push_args is the code to do that.
950 -- d_after_args is the stack depth once the args are on.
952 -- Get the result rep.
953 (returns_void, r_rep)
954 = case maybe_getCCallReturnRep (idType fn) of
955 Nothing -> (True, VoidArg)
956 Just rr -> (False, rr)
958 Because the Haskell stack grows down, the a_reps refer to
959 lowest to highest addresses in that order. The args for the call
960 are on the stack. Now push an unboxed Addr# indicating
961 the C function to call. Then push a dummy placeholder for the
962 result. Finally, emit a CCALL insn with an offset pointing to the
963 Addr# just pushed, and a literal field holding the mallocville
964 address of the piece of marshalling code we generate.
965 So, just prior to the CCALL insn, the stack looks like this
966 (growing down, as usual):
971 Addr# address_of_C_fn
972 <placeholder-for-result#> (must be an unboxed type)
974 The interpreter then calls the marshall code mentioned
975 in the CCALL insn, passing it (& <placeholder-for-result#>),
976 that is, the addr of the topmost word in the stack.
977 When this returns, the placeholder will have been
978 filled in. The placeholder is slid down to the sequel
979 depth, and we RETURN.
981 This arrangement makes it simple to do f-i-dynamic since the Addr#
982 value is the first arg anyway.
984 The marshalling code is generated specifically for this
985 call site, and so knows exactly the (Haskell) stack
986 offsets of the args, fn address and placeholder. It
987 copies the args to the C stack, calls the stacked addr,
988 and parks the result back in the placeholder. The interpreter
989 calls it as a normal C call, assuming it has a signature
990 void marshall_code ( StgWord* ptr_to_top_of_stack )
992 -- resolve static address
996 -> return (False, panic "ByteCodeGen.generateCCall(dyn)")
998 -> do res <- ioToBc (lookupStaticPtr target)
1001 (is_static, static_target_addr) <- get_target_info
1004 -- Get the arg reps, zapping the leading Addr# in the dynamic case
1005 a_reps -- | trace (showSDoc (ppr a_reps_pushed_RAW)) False = error "???"
1006 | is_static = a_reps_pushed_RAW
1007 | otherwise = if null a_reps_pushed_RAW
1008 then panic "ByteCodeGen.generateCCall: dyn with no args"
1009 else tail a_reps_pushed_RAW
1012 (push_Addr, d_after_Addr)
1014 = (toOL [PUSH_UBX (Right static_target_addr) addr_sizeW],
1015 d_after_args + addr_sizeW)
1016 | otherwise -- is already on the stack
1017 = (nilOL, d_after_args)
1019 -- Push the return placeholder. For a call returning nothing,
1020 -- this is a VoidArg (tag).
1021 r_sizeW = cgRepSizeW r_rep
1022 d_after_r = d_after_Addr + r_sizeW
1023 r_lit = mkDummyLiteral r_rep
1024 push_r = (if returns_void
1026 else unitOL (PUSH_UBX (Left r_lit) r_sizeW))
1028 -- generate the marshalling code we're going to call
1031 arg1_offW = r_sizeW + addr_sizeW
1032 args_offW = map (arg1_offW +)
1033 (init (scanl (+) 0 (map cgRepSizeW a_reps)))
1035 addr_of_marshaller <- ioToBc (mkMarshalCode cconv
1036 (r_offW, r_rep) addr_offW
1037 (zip args_offW a_reps))
1038 recordItblMallocBc (ItblPtr (castFunPtrToPtr addr_of_marshaller))
1040 -- Offset of the next stack frame down the stack. The CCALL
1041 -- instruction needs to describe the chunk of stack containing
1042 -- the ccall args to the GC, so it needs to know how large it
1043 -- is. See comment in Interpreter.c with the CCALL instruction.
1044 stk_offset = d_after_r - s
1047 do_call = unitOL (CCALL stk_offset (castFunPtrToPtr addr_of_marshaller))
1049 wrapup = mkSLIDE r_sizeW (d_after_r - r_sizeW - s)
1050 `snocOL` RETURN_UBX r_rep
1052 --trace (show (arg1_offW, args_offW , (map cgRepSizeW a_reps) )) $
1055 push_Addr `appOL` push_r `appOL` do_call `appOL` wrapup
1059 -- Make a dummy literal, to be used as a placeholder for FFI return
1060 -- values on the stack.
1061 mkDummyLiteral :: CgRep -> Literal
1064 NonPtrArg -> MachWord 0
1065 DoubleArg -> MachDouble 0
1066 FloatArg -> MachFloat 0
1067 LongArg -> MachWord64 0
1068 _ -> moan64 "mkDummyLiteral" (ppr pr)
1072 -- GHC.Prim.Char# -> GHC.Prim.State# GHC.Prim.RealWorld
1073 -- -> (# GHC.Prim.State# GHC.Prim.RealWorld, GHC.Prim.Int# #)
1076 -- and check that an unboxed pair is returned wherein the first arg is VoidArg'd.
1078 -- Alternatively, for call-targets returning nothing, convert
1080 -- GHC.Prim.Char# -> GHC.Prim.State# GHC.Prim.RealWorld
1081 -- -> (# GHC.Prim.State# GHC.Prim.RealWorld #)
1085 maybe_getCCallReturnRep :: Type -> Maybe CgRep
1086 maybe_getCCallReturnRep fn_ty
1087 = let (a_tys, r_ty) = splitFunTys (dropForAlls fn_ty)
1089 = if isSingleton r_reps then Nothing else Just (r_reps !! 1)
1091 = case splitTyConApp_maybe (repType r_ty) of
1092 (Just (tyc, tys)) -> (tyc, map typeCgRep tys)
1094 ok = ( ( r_reps `lengthIs` 2 && VoidArg == head r_reps)
1095 || r_reps == [VoidArg] )
1096 && isUnboxedTupleTyCon r_tycon
1097 && case maybe_r_rep_to_go of
1099 Just r_rep -> r_rep /= PtrArg
1100 -- if it was, it would be impossible
1101 -- to create a valid return value
1102 -- placeholder on the stack
1103 blargh = pprPanic "maybe_getCCallReturn: can't handle:"
1106 --trace (showSDoc (ppr (a_reps, r_reps))) $
1107 if ok then maybe_r_rep_to_go else blargh
1109 -- Compile code which expects an unboxed Int on the top of stack,
1110 -- (call it i), and pushes the i'th closure in the supplied list
1111 -- as a consequence.
1112 implement_tagToId :: [Name] -> BcM BCInstrList
1113 implement_tagToId names
1114 = ASSERT( notNull names )
1115 do labels <- getLabelsBc (length names)
1116 label_fail <- getLabelBc
1117 label_exit <- getLabelBc
1118 let infos = zip4 labels (tail labels ++ [label_fail])
1120 steps = map (mkStep label_exit) infos
1121 return (concatOL steps
1123 toOL [LABEL label_fail, CASEFAIL, LABEL label_exit])
1125 mkStep l_exit (my_label, next_label, n, name_for_n)
1126 = toOL [LABEL my_label,
1127 TESTEQ_I n next_label,
1132 -- -----------------------------------------------------------------------------
1135 -- Push an atom onto the stack, returning suitable code & number of
1136 -- stack words used.
1138 -- The env p must map each variable to the highest- numbered stack
1139 -- slot for it. For example, if the stack has depth 4 and we
1140 -- tagged-ly push (v :: Int#) on it, the value will be in stack[4],
1141 -- the tag in stack[5], the stack will have depth 6, and p must map v
1142 -- to 5 and not to 4. Stack locations are numbered from zero, so a
1143 -- depth 6 stack has valid words 0 .. 5.
1145 pushAtom :: Int -> BCEnv -> AnnExpr' Id VarSet -> BcM (BCInstrList, Int)
1147 pushAtom d p (AnnApp f (_, AnnType _))
1148 = pushAtom d p (snd f)
1150 pushAtom d p (AnnNote note e)
1151 = pushAtom d p (snd e)
1153 pushAtom d p (AnnLam x e)
1155 = pushAtom d p (snd e)
1157 pushAtom d p (AnnVar v)
1159 | idCgRep v == VoidArg
1163 = pprPanic "pushAtom: shouldn't get an FCallId here" (ppr v)
1165 | Just primop <- isPrimOpId_maybe v
1166 = return (unitOL (PUSH_PRIMOP primop), 1)
1168 | Just d_v <- lookupBCEnv_maybe p v -- v is a local variable
1169 = return (toOL (nOfThem sz (PUSH_L (d-d_v+sz-2))), sz)
1170 -- d - d_v the number of words between the TOS
1171 -- and the 1st slot of the object
1173 -- d - d_v - 1 the offset from the TOS of the 1st slot
1175 -- d - d_v - 1 + sz - 1 the offset from the TOS of the last slot
1178 -- Having found the last slot, we proceed to copy the right number of
1179 -- slots on to the top of the stack.
1181 | otherwise -- v must be a global variable
1183 return (unitOL (PUSH_G (getName v)), sz)
1189 pushAtom d p (AnnLit lit)
1191 MachLabel fs _ -> code NonPtrArg
1192 MachWord w -> code NonPtrArg
1193 MachInt i -> code PtrArg
1194 MachFloat r -> code FloatArg
1195 MachDouble r -> code DoubleArg
1196 MachChar c -> code NonPtrArg
1197 MachStr s -> pushStr s
1200 = let size_host_words = cgRepSizeW rep
1201 in return (unitOL (PUSH_UBX (Left lit) size_host_words),
1205 = let getMallocvilleAddr
1207 FastString _ n _ fp _ ->
1208 -- we could grab the Ptr from the ForeignPtr,
1209 -- but then we have no way to control its lifetime.
1210 -- In reality it'll probably stay alive long enoungh
1211 -- by virtue of the global FastString table, but
1212 -- to be on the safe side we copy the string into
1213 -- a malloc'd area of memory.
1214 do ptr <- ioToBc (mallocBytes (n+1))
1217 withForeignPtr fp $ \p -> do
1218 memcpy ptr p (fromIntegral n)
1219 pokeByteOff ptr n (fromIntegral (ord '\0') :: Word8)
1223 addr <- getMallocvilleAddr
1224 -- Get the addr on the stack, untaggedly
1225 return (unitOL (PUSH_UBX (Right addr) 1), 1)
1227 pushAtom d p (AnnCast e _)
1228 = pushAtom d p (snd e)
1231 = pprPanic "ByteCodeGen.pushAtom"
1232 (pprCoreExpr (deAnnotate (undefined, other)))
1234 foreign import ccall unsafe "memcpy"
1235 memcpy :: Ptr a -> Ptr b -> CSize -> IO ()
1238 -- -----------------------------------------------------------------------------
1239 -- Given a bunch of alts code and their discrs, do the donkey work
1240 -- of making a multiway branch using a switch tree.
1241 -- What a load of hassle!
1243 mkMultiBranch :: Maybe Int -- # datacons in tycon, if alg alt
1244 -- a hint; generates better code
1245 -- Nothing is always safe
1246 -> [(Discr, BCInstrList)]
1248 mkMultiBranch maybe_ncons raw_ways
1249 = let d_way = filter (isNoDiscr.fst) raw_ways
1251 (\w1 w2 -> leAlt (fst w1) (fst w2))
1252 (filter (not.isNoDiscr.fst) raw_ways)
1254 mkTree :: [(Discr, BCInstrList)] -> Discr -> Discr -> BcM BCInstrList
1255 mkTree [] range_lo range_hi = return the_default
1257 mkTree [val] range_lo range_hi
1258 | range_lo `eqAlt` range_hi
1261 = do label_neq <- getLabelBc
1262 return (mkTestEQ (fst val) label_neq
1264 `appOL` unitOL (LABEL label_neq)
1265 `appOL` the_default))
1267 mkTree vals range_lo range_hi
1268 = let n = length vals `div` 2
1269 vals_lo = take n vals
1270 vals_hi = drop n vals
1271 v_mid = fst (head vals_hi)
1273 label_geq <- getLabelBc
1274 code_lo <- mkTree vals_lo range_lo (dec v_mid)
1275 code_hi <- mkTree vals_hi v_mid range_hi
1276 return (mkTestLT v_mid label_geq
1278 `appOL` unitOL (LABEL label_geq)
1282 = case d_way of [] -> unitOL CASEFAIL
1285 -- None of these will be needed if there are no non-default alts
1286 (mkTestLT, mkTestEQ, init_lo, init_hi)
1288 = panic "mkMultiBranch: awesome foursome"
1290 = case fst (head notd_ways) of {
1291 DiscrI _ -> ( \(DiscrI i) fail_label -> TESTLT_I i fail_label,
1292 \(DiscrI i) fail_label -> TESTEQ_I i fail_label,
1295 DiscrF _ -> ( \(DiscrF f) fail_label -> TESTLT_F f fail_label,
1296 \(DiscrF f) fail_label -> TESTEQ_F f fail_label,
1299 DiscrD _ -> ( \(DiscrD d) fail_label -> TESTLT_D d fail_label,
1300 \(DiscrD d) fail_label -> TESTEQ_D d fail_label,
1303 DiscrP _ -> ( \(DiscrP i) fail_label -> TESTLT_P i fail_label,
1304 \(DiscrP i) fail_label -> TESTEQ_P i fail_label,
1306 DiscrP algMaxBound )
1309 (algMinBound, algMaxBound)
1310 = case maybe_ncons of
1311 Just n -> (0, n - 1)
1312 Nothing -> (minBound, maxBound)
1314 (DiscrI i1) `eqAlt` (DiscrI i2) = i1 == i2
1315 (DiscrF f1) `eqAlt` (DiscrF f2) = f1 == f2
1316 (DiscrD d1) `eqAlt` (DiscrD d2) = d1 == d2
1317 (DiscrP i1) `eqAlt` (DiscrP i2) = i1 == i2
1318 NoDiscr `eqAlt` NoDiscr = True
1321 (DiscrI i1) `leAlt` (DiscrI i2) = i1 <= i2
1322 (DiscrF f1) `leAlt` (DiscrF f2) = f1 <= f2
1323 (DiscrD d1) `leAlt` (DiscrD d2) = d1 <= d2
1324 (DiscrP i1) `leAlt` (DiscrP i2) = i1 <= i2
1325 NoDiscr `leAlt` NoDiscr = True
1328 isNoDiscr NoDiscr = True
1331 dec (DiscrI i) = DiscrI (i-1)
1332 dec (DiscrP i) = DiscrP (i-1)
1333 dec other = other -- not really right, but if you
1334 -- do cases on floating values, you'll get what you deserve
1336 -- same snotty comment applies to the following
1338 minD, maxD :: Double
1344 mkTree notd_ways init_lo init_hi
1347 -- -----------------------------------------------------------------------------
1348 -- Supporting junk for the compilation schemes
1350 -- Describes case alts
1358 instance Outputable Discr where
1359 ppr (DiscrI i) = int i
1360 ppr (DiscrF f) = text (show f)
1361 ppr (DiscrD d) = text (show d)
1362 ppr (DiscrP i) = int i
1363 ppr NoDiscr = text "DEF"
1366 lookupBCEnv_maybe :: BCEnv -> Id -> Maybe Int
1367 lookupBCEnv_maybe = lookupFM
1369 idSizeW :: Id -> Int
1370 idSizeW id = cgRepSizeW (typeCgRep (idType id))
1372 unboxedTupleException :: a
1373 unboxedTupleException
1376 ("Bytecode generator can't handle unboxed tuples. Possibly due\n" ++
1377 "\tto foreign import/export decls in source. Workaround:\n" ++
1378 "\tcompile this module to a .o file, then restart session."))
1381 mkSLIDE n d = if d == 0 then nilOL else unitOL (SLIDE n d)
1384 splitApp :: AnnExpr' id ann -> (AnnExpr' id ann, [AnnExpr' id ann])
1385 -- The arguments are returned in *right-to-left* order
1386 splitApp (AnnApp (_,f) (_,a))
1387 | isTypeAtom a = splitApp f
1388 | otherwise = case splitApp f of
1389 (f', as) -> (f', a:as)
1390 splitApp (AnnNote n (_,e)) = splitApp e
1391 splitApp (AnnCast (_,e) _) = splitApp e
1392 splitApp e = (e, [])
1395 isTypeAtom :: AnnExpr' id ann -> Bool
1396 isTypeAtom (AnnType _) = True
1397 isTypeAtom _ = False
1399 isVoidArgAtom :: AnnExpr' id ann -> Bool
1400 isVoidArgAtom (AnnVar v) = typeCgRep (idType v) == VoidArg
1401 isVoidArgAtom (AnnNote n (_,e)) = isVoidArgAtom e
1402 isVoidArgAtom (AnnCast (_,e) _) = isVoidArgAtom e
1403 isVoidArgAtom _ = False
1405 atomRep :: AnnExpr' Id ann -> CgRep
1406 atomRep (AnnVar v) = typeCgRep (idType v)
1407 atomRep (AnnLit l) = typeCgRep (literalType l)
1408 atomRep (AnnNote n b) = atomRep (snd b)
1409 atomRep (AnnApp f (_, AnnType _)) = atomRep (snd f)
1410 atomRep (AnnLam x e) | isTyVar x = atomRep (snd e)
1411 atomRep (AnnCast b _) = atomRep (snd b)
1412 atomRep other = pprPanic "atomRep" (ppr (deAnnotate (undefined,other)))
1414 isPtrAtom :: AnnExpr' Id ann -> Bool
1415 isPtrAtom e = atomRep e == PtrArg
1417 -- Let szsw be the sizes in words of some items pushed onto the stack,
1418 -- which has initial depth d'. Return the values which the stack environment
1419 -- should map these items to.
1420 mkStackOffsets :: Int -> [Int] -> [Int]
1421 mkStackOffsets original_depth szsw
1422 = map (subtract 1) (tail (scanl (+) original_depth szsw))
1424 -- -----------------------------------------------------------------------------
1425 -- The bytecode generator's monad
1427 type BcPtr = Either ItblPtr (Ptr ())
1431 uniqSupply :: UniqSupply, -- for generating fresh variable names
1432 nextlabel :: Int, -- for generating local labels
1433 malloced :: [BcPtr], -- thunks malloced for current BCO
1434 -- Should be free()d when it is GCd
1435 breakArray :: BreakArray -- array of breakpoint flags
1438 newtype BcM r = BcM (BcM_State -> IO (BcM_State, r))
1440 ioToBc :: IO a -> BcM a
1441 ioToBc io = BcM $ \st -> do
1445 runBc :: UniqSupply -> ModBreaks -> BcM r -> IO (BcM_State, r)
1446 runBc us modBreaks (BcM m)
1447 = m (BcM_State us 0 [] breakArray)
1449 breakArray = modBreaks_flags modBreaks
1451 thenBc :: BcM a -> (a -> BcM b) -> BcM b
1452 thenBc (BcM expr) cont = BcM $ \st0 -> do
1453 (st1, q) <- expr st0
1458 thenBc_ :: BcM a -> BcM b -> BcM b
1459 thenBc_ (BcM expr) (BcM cont) = BcM $ \st0 -> do
1460 (st1, q) <- expr st0
1461 (st2, r) <- cont st1
1464 returnBc :: a -> BcM a
1465 returnBc result = BcM $ \st -> (return (st, result))
1467 instance Monad BcM where
1472 emitBc :: ([BcPtr] -> ProtoBCO Name) -> BcM (ProtoBCO Name)
1474 = BcM $ \st -> return (st{malloced=[]}, bco (malloced st))
1476 recordMallocBc :: Ptr a -> BcM ()
1478 = BcM $ \st -> return (st{malloced = Right (castPtr a) : malloced st}, ())
1480 recordItblMallocBc :: ItblPtr -> BcM ()
1481 recordItblMallocBc a
1482 = BcM $ \st -> return (st{malloced = Left a : malloced st}, ())
1484 getLabelBc :: BcM Int
1486 = BcM $ \st -> return (st{nextlabel = 1 + nextlabel st}, nextlabel st)
1488 getLabelsBc :: Int -> BcM [Int]
1490 = BcM $ \st -> let ctr = nextlabel st
1491 in return (st{nextlabel = ctr+n}, [ctr .. ctr+n-1])
1493 getBreakArray :: BcM BreakArray
1494 getBreakArray = BcM $ \st -> return (st, breakArray st)
1496 newUnique :: BcM Unique
1498 \st -> case splitUniqSupply (uniqSupply st) of
1499 (us1, us2) -> let newState = st { uniqSupply = us2 }
1500 in return (newState, uniqFromSupply us1)
1502 newId :: Type -> BcM Id
1505 return $ mkSysLocal FSLIT("ticked") uniq ty